README.md

SubT

The tools for SubT interestingness dataset released by The Air Lab at Carnegie Mellon University.

Description

Refer to The AirLab Dataset Instructions

This repo contains the following tools:

• PyTorch dataloader (No need to install PyTorch, if you don't need this function).

• Python evaluation tools for the daseset.

  python -m pip install --user numpy scipy matplotlib ipython jupyter pandas sympy nose
  

• ROS bag launchfiles and visualization (No need to install ROS, if you don't need this function).

PyTorch dataloader

Usage example:

    from subt import SubT
    import torch.utils.data as Data
    import torchvision.transforms as transforms

    transform = transforms.Compose([transforms.Resize((320,320)), transforms.ToTensor()])
    data = SubT(root='/data/location', train=False, transform=transform, test_data=1)
    loader = Data.DataLoader(dataset=data, batch_size=1, shuffle=False)

The dataloader is only for SubT front camera data (SubTF).

Evaluation Tools

For definition of the evaluation metric, checkout our paper.

Usage Example

You have to provide your result files.

Result example files are provided in folder 'results-example'.

For single sequence:

  python script/evaluation.py --source 'ground-truth/0817-ugv0-tunnel0-interest-1.txt' --target 'results-example/SubTF-0-2020-03-04-21:43:27-example.txt'
  # source: ground-truth file
  # target: results file

For overall performance on entire dataset:

  python script/performance.py --save-flag example --root results-example --category interest-1
  # root: location of result files
  # save-flag: flag of result files
  #     You have to name the result files following the formate of "SubTF-5-2020-03-04-22:06:43-example.txt", 
  #     where only the sequence ID, date, and time can be different.
  # category: interest-1 or interest-2

For example, after you run the following overall performance script.

  python script/performance.py --save-flag example --root results-example --category interest-2 --delta 1 2 4

You will get:

• Mean Accuracy for delta = [1,2,4], respectively.

  mean accuracy: [0.26001525 0.37333027 0.5226512 ]
  

• One Text file located in 'performance' folder.

  It contains data of AUC-OP (three curves for delta = [1,2,4])
  

• One overal performance figure, seven single sequence figures, e.g.

  

ROS bag launchfile and visualization

To use this function, you have to put this repo in your ROS workplace and compile, e.g.,

  cd ~/catkin_ws/src
  git clone https://github.com/wang-chen/SubT.git
  cd ~/catkin_ws && catkin_make

Then run the launch file

 roslaunch SubT subt.launch

Note: you need to specify the argument 'datalocation' in 'subt.larunch' to find your bag data.

 <arg name="datalocation" default="/data/datasets"/>

You also need to specify the sequence ID, e.g. SubT0.

 <node pkg="rosbag" type="play" name="rosbag" args="--clock -r 3 $(arg SubT0)"/>

• The corresponding sequence ID is listed:

  Data sequence	ID	ROS args
  0817-ugv0-tunnel0	0	SubT0
  0817-ugv1-tunnel0	1	SubT1
  0818-ugv0-tunnel1	2	SubT2
  0818-ugv1-tunnel1	3	SubT3
  0820-ugv0-tunnel1	4	SubT4
  0821-ugv0-tunnel0	5	SubT5
  0821-ugv1-tunnel0	6	SubT6

Then you'll get:

Citation

  @article{wang2020visual,
     title={Visual Memorability for Robotic Interestingness via Unsupervised Online Learning},
     author={Wang, Chen and Wang, Wenshan and Qiu, Yuheng and Hu, Yafei and Scherer, Sebastian},
     journal={arXiv preprint arXiv:2005.08829},
     year={2020}
     }

Download this paper